magiswarm

Inspiration

Justin is a researcher currently working as an RA without any funding. Recently, administrative policy has brought research grants to a halt and it is increasingly difficult to do viable research with enough hands on board.

What it does

Our project Magi Swarm (Inspired by "MAGI" from Neon Genesis Evangelion) uses a combination of custom AI and productivity tools to form fast and powerful research labs at the tips of your fingers. We integrate a swarm of custom specialized AI agents acting as research aides with tools like knowledge graphs, a paper library, and a simulated "meeting room."

How we built it

We used Modal to host dynamically allocated any number of Qwen2.5 agents that fan out to GPUs as needed and utilize Modal Sandboxes for iterative and responsive development. We also use Modal for API hosting to handle end to end communication and tool use. We use supermemory to handle long context retrieval and forming the knowledge graph through semantic connectivity. We use ElevenLabs TTS to provide natural language meeting notes that summarize each agents actions. We store all of our logs and data on Supabase and deploy using Vercel.

Challenges we ran into

Since our app is so feature diverse, each development step required a different set of skills and required us to shift our focus back and forth quickly.

Accomplishments that we're proud of

We are proud of how much we were able to develop for the time we were given with only two of us. We believe that we have made a genuinely useful and impactful tool to solve a problem important to us.

What we learned

We learned what it means to push a product that builds upon a number of technologies and how to effectively unify them in a way that benefits each component.

What's next for magiswarm

There is a lot of efficiency optimizations and prompt engineering/model choice that can be done to better define each agents task and allow them to execute them in a more effective manner. We would also like to increase agent collaboration and allow for a less strict environment for them to ideate in.